Results of training on 3D data

i did some reimplementation of benchmark(best model) Rsconv on my own dataset. Normally, other algorithm perform well, at start the train and test accuracy both surges and be steady at certain epoch.

However, when i train this model. after i adjust learning rate the performance of model is till around 60.

I also reimplement it on benchmark dataset modelnet10, it works well.
I guess the reason is that the Rsconv designed for large dataset?